how does the validation split work when calling fit? - neural-network

When I specify the validation_split parameter when calling the fit method of the model class, is the same validation data used at every epoch? Or does the validation data change for every epoch?

It uses the same validation data for every epoch.
If it didn't then it wouldn't be validation data as the model would learn have seen elements in the validation set during previous epochs and and thus would have the possibility of overfitting to them.

Related

How to change time frame of the model?

I’m making a model which initially could be a minute based model, but now I have to change the time frame. Is it possible to change time frame from minutes to days?
it is advised not to change the model time units once you created the model since there are a few functions that cannot be adapted (such as toTimeoutInCalendar), it is also advised to use the time units in all functions so you CAN change the time units early on, for example time(MINUTE)
with that being said, you can change it
Adding to Felipe's correct answer: there is no need ever to change the model time unit, if you use calls of time() and date() correctly:
As long as you always specify the time unit in the calls (time(SECOND) or date(YEAR)), your model will not suffer changing the time unit.
So you can "refactor" your model by specifying all such calls correctly and then, changing the model time unit is fine. It will not even have an impact on the model performance our results!
Check the help to understand these concepts, maybe easier than redoing the entire model.
PS: Even without it, there is literally no need to change the model time unit. Just adjust the run speed of the simulation experiment if it was too slow/fast

Literature for Classification Problem with Changing Classes

I am currently looking at a text classification problem (say N classes), for which labeled training data exists. Now, ocasionally, a new class is created and some of the labels in the "old" training data become wrong because they now should have the new class label. So the new class recruits from the old classes.
We can assume that we have some new labeled data for the new class, or even that from an input stream of new data we eventually obtain the correct labels by human verification (the goal, however, is to require as few manual corrections as possible).
How to set up a classifier that may face new "recruiting" classes from time to time? Are you aware of approaches/literature for the specific setting described above?
Perhaps, basic strategies may include
trying to relabel the training data and re-train,
using incremental classifiers (e.g., KNN)

Best way to save a data structure with date-related varying attributes in Swift/CoreData

I am thinking about how to best represent a certain data structure efficiently in Swift using CoreData. What I need is work with accounts (like savings, earning etc.). So what would probably make sense is an account class, where each account instance might have multiple characteristics like e.g. ACCOUNT_TYPE_ID which do not change. The core however is the VALUE attribute, which would hold the value of the account at a certain point in time. The complex thing here is that this value obviously might change over time (lets say on a daily basis, abstracting for intra-day changes) and I would need to be able to get the value of each instance for any given date. E.g. I might have my savings_private for which I would want to get the value at each month-end. This value might have changed, but as well could stay the same for various days/months. How could this most efficiently (when it comes to used space but - and that is even more important - be able to access computationally efficient) be done with a CoreData Entity/Class? I was thinking about maybe always starting with zero and then only somehow save the changes plus the date of change and then have some method for the call which would add all changes up to a date parameter - but was curious about what a best-practice might be here, as I guess I am not the first one trying to solve this.

Access the simulation parameter in Modelica

I have a model in my Modelica and I use Dymola to compile this model. In my model I need the simulation information "Output Interval length". I have searched for it but I could not get the useful information. Is there any other possible way we could access simulation information.
If you are simply trying to get the results reported at specific intervals, you can use a sample operator to achieve that. That would force the solution to be computed at specific times without directly specifying something like the time step.
The important point to understand here is that a model where the behavior of the model depends on the numerical integration is highly suspect and I've never seen a case where the behavior couldn't be described without knowledge of the solution method. Said another way, "mother nature" doesn't know anything about "time steps". :-)
You could use a clocked system with an integrator.
For an Example, see File -->Libraries-->Modelica_Synchronous --> Examples --> Systems --> Controlled_mixing_unit in Dymola
There the period (i.e. in this case the timestep of the explicit Euler method) is a parameter of the periodic clock)
Modelica by design prohibits accessing any numerical solver internals, so you cannot access it. The output interval length also cannot be determined by the model in any reliable way since the solver will take internal steps longer than the output interval and then interpolate values for the result file.
You could create a function that reads the dsin.txt file and extracts that information.

How to accumulate errors in a functional way upon validating database object?

I have Product case class, which is returned by DAO layer (using Salat). User who is creating a product first time status of the product remains as "draft" where no field (of product) is mandatory.
What are the best functional ways to validate 10 of product's attributes, accumulate all validation errors into a single entity and then pass all errors at once in a JSON format to front end?
I assume the core of the question is how to accumulate errors--JSON formatting is a separate issue and does not depend upon how you have collected your errors.
If it's really just a validation issue, you can have a series of methods
def problemWithX: Option[String] = ...
which return Some(errorMessage) if they are invalid or None if they're okay. Then it's as simple as
List(problemWithX, problemWithY, ...).flatten
to create a list of all of your errors. If your list is empty, you're good to go. If not, you have the errors listed. Creating some sensible error report is the job of the problemWithX method--and of course you need to decide whether merely a string or more complex information is necessary. (You may even need to define a Invalid trait and have classes extend it to handle different conditions.)
This is exactly what ScalaZ's Validation type is for.